Jitter in PCIe application on embedded boards with PLL Zero delay Clock buffer

Transcription

1 Jitter in PCIe application on embedded boards with PLL Zero delay Clock buffer Hermann Ruckerbauer EKH - EyeKnowHow Deggendorf, Germany Hermann.Ruckerbauer@EyeKnowHow.de

2 Agenda 1) PCI-Express Clocking 2) Motivation and Background 2a) Basics 2b) Clocking in different PCIe generations 2c) Different PCIe clocking architectures 3) Clock Compliance Test 4) Conclusion EKH - EyeKnowHow 3/1/2014 2

3 PCIe Clocking PCIe are working at speeds up to 8Gb/s the RefClk only works at 100MHz! While the data signals are treated as high speed, the clock is often treated as low speed signal But the PCIe specification relies on a well defined Jitter behavior, so it is very important to verify the RefClk behavior very carefully PCIe RefClk needs also be measured as part of the compliance measurements in a PCIe system! Jitter as well as Signal Integrity and AC parameter definition needs to be verified in these tests! EKH - EyeKnowHow 3/1/2014 3

4 Agenda 1) PCI-Express Clocking 2) Motivation and Background 2a) Basics 2b) Clocking in different PCIe generations 2c) Different PCIe clocking architectures 3) Clock Compliance Test 4) Conclusion EKH - EyeKnowHow 3/1/2014 4

5 Important Basics Jitter amplitude Jitter Explanation out of Agilent Jitter seminar 2006 Jitter Analysis Techniques for High Data Rates (Application Note 1432) Clock at ideal position EKH - EyeKnowHow 3/1/2014 5

6 Important Basics PLL Jitter Transfer Function PLL should follow Low Frequency Jitter (e. g. SSC or Wander) and cancel High Frequency Jitter Jitter Transfer function of a PLL Ideal behavior (blue line) As long PLL follows low frequency Jitter in the eye this is not seen in the Eye Diagram Possible real behavior (red Dots) If the PLL/CDR can not follow this Jitter it will show up in the Eye Diagram Things like Jitter Peaking are not considered in this picture 3 MHz SSC (30kHz) Effect of CDR (and Jitter Transfer function) on Jitter EKH - EyeKnowHow 3/1/2014 6

7 Important Basics Order of a PLL 1 st or 2 nd Order of a PLL*? *Source: Tutorial 18 and 19 EKH - EyeKnowHow 3/1/2014 7

8 Important Basics Order of a PLL for measurements The additional filter in a 2 nd order PLL would clean up the jitter in this signal If a system uses such a PLL for CDR it would not see this jitter PCIe devices should use at least 2 nd order PLLs 1 st order PLL 2 nd order PLL EKH - EyeKnowHow 3/1/2014 8

9 Agenda 1) PCI-Express Clocking 2) Motivation and Background 2a) Basics 2b) Clocking in different PCIe generations 2c) Different PCIe clocking architectures 3) Clock Compliance Test 4) Conclusion EKH - EyeKnowHow 3/1/2014 9

10 Clock Specification over PCIe Generations First major impact on clock performance was the change from PCIe 1.0a to 1.1: The PCI Express 1.0a specification failed to specify the input bandwidth the reference clock receiver or phase jitter of the reference clock itself. This is important because jitter that lies within the loop bandwidth the receiver PLL for the reference clock will transfer onto the high speed data lines. This hole in the PCI Express specification was corrected in the 1.1 update * In Gen1 the RefClk spec was part of the CEM spec, in Gen2 it moved to the base spec Other parameters are added Gen3 added again some parameters that are required to be measured for compliance testing *Source: Agilent application note EN EKH - EyeKnowHow 3/1/

11 Clock Specification over PCIe Generations Gen2 Compliance testing changed to Dual Port Testing. This test method convolves Signal and clock traces and calculate Signal quality for data traces Dual Port Signal compliance testing does not mean separate clock tests are obsolete! Lower speed RefClk specs are no subsets from Gen3 specification. Each generation covers it s own frequency requirement RefClk Tests are required for each generation separately! EKH - EyeKnowHow 3/1/

12 Agenda 1) PCI-Express Clocking 2) Motivation and Background 2a) Basics 2b) Clocking in different PCIe generations 2c) Different PCIe clocking architectures 3) Clock Compliance Test 4) Conclusion EKH - EyeKnowHow 3/1/

13 Clocking Architectures Three different clocking architectures are defined: Common clocked architecture Data clocked architecture Separate clocked architecture This would require much tighter clock spec and no SSC in the system. This configuration is not covered in this presentation Even called clocking architectures these are receiver (RX) implementation architectures Usually a system designer does not know about the RX clocking architecture of the used devices (e. g. for AddIn Cards). So from a system perspective it is required to support both, common clocked and data clocked architecture! EKH - EyeKnowHow 3/1/

14 Example: Common RefClk architecture Which Transfer is shown? Host to AddIn Card? Which components are where? How to calculate the transport delay T2 T1? TX Latch TX PLL T2 RefClk T1 RX Latch RX CDR RX PLL How does this map to embedded Applications? EKH - EyeKnowHow 3/1/

15 Common clocked RX architecture Transfer from ComExpress Assuming Zero Delay for the clock buffer there is only small delta between RefClk and TX Transport Delay ComExpress Board PCIe AddIn Card RX Latch RX CDR TX/RX PLL Transport Delay RefCLK ZD Clock Buffer TX Latch RX/TX PLL TX Latch Transport Delay TX RX CDR RX Latch Carrier Board Jitter Domain on RX is the same for Clk and Data! Small delta for transport delay due to system configuration EKH - EyeKnowHow 3/1/

16 Data clocked RX architecture Transfer from ComExpress No Transport Delay requirement! ComExpress Board RX Latch RX CDR PLL not used for RX Data capture PCIe AddIn Card TX Latch TX/RX PLL ZD Clock Buffer TX PLL TX Latch RX CDR RX Latch Carrier Board EKH - EyeKnowHow 3/1/

17 Data clocked RX architecture Transfer from AddIn Card No Transport delay requirement ComExpress Board PCIe AddIn Card RX Latch RX CDR TX Latch TX PLL TX PLL ZD Clock Buffer Carrier Board PLL not used for RX Data capture EKH - EyeKnowHow 3/1/

18 Transport Delay RX RefClk Common clocked RX architecture Transfer from AddIn Card Transport delay Delta requirement < 12ns ComExpress Board Transport Delay TX_RefCLK + TX PCIe AddIn Card RX Latch RX CDR TX Latch TX PLL TX/RX PLL ZD Clock Buffer Carrier Board RX Latch AddIn Card TX Latch T2 T1 RX CDR Host TX PLL RefClk RX PLL EKH - EyeKnowHow 3/1/

19 Common Clocked RX architecture Transport delay calculation Transport delay calculation for embedded systems Parameters to be considered: Signal flight times on PCB CPU Module Carrier Board AddIn Card Delays introduced from connectors RX/TX device delays Zero Delay Buffer EKH - EyeKnowHow 3/1/

20 Common Clocked RX architecture Transport delay calculation Signal flight time on PCB is around 175ps/inch 4inch routing on AddIn Card for clock and data calculates to: 2signals x 4inch x 175ps = 1.4ns Considering a connector delay with 150ps and knowing that each signal crosses two connectors calculates to 0.6ns Zero Delay clock buffer need to be considered with several parameters e. g. clock output skew. Reference [4] calculates these to another 2ns TX internal delays need also be considered with 2ns These delays take 6ns from the specified 12ns maximum transport delay EKH - EyeKnowHow 3/1/

21 Common Clocked RX architecture Transport delay calculation Remaining 6ns need to be distributed to clock and data lines, so 3ns for each signal With 175ps/inch this allows 17Inch routing on carrier board and CPU Module 5Inch module routing leaves 12 inch routing on carrier board The special implementation with Connectors and Zero Delay clock Buffer limits the solution space for PCIe clock distribution to around 12 Inch clock routing! EKH - EyeKnowHow 3/1/

22 Agenda 1) PCI-Express Clocking 2) Motivation and Background 2a) Basics 2b) Clocking in different PCIe generations 2c) Different PCIe clocking architectures 3) Clock Compliance Test 4) Conclusion EKH - EyeKnowHow 3/1/

23 Clock Compliance: Test Setup Specified clock compliance test Setup No Termination and 2pF load to GND Test Point RefClk Generator DUT 12inch 100Ohm Differential PCB Trace 2pF load 2pF load RefClk Test Setup EKH - EyeKnowHow 3/1/

24 Clock Compliance: Test Setup The CLB (Compliance Load Board) is not optimized for clock compliance tests Normal test configuration connect 50Ohm SMA cables to Scope input with 50 Ohm termination This does not fit to the specified test load configuration: No Termination and 2pF load to GND Connect for differential active Probe Usually 50 Ohm Termination Edge SMP 2pF load EKH - EyeKnowHow 3/1/

25 Clock Compliance: Test Setup Difficult to create a common setup for all measurements. Standard configuration in most cases will be SMP/SMA cables to 50Ohm scope input Due to 50 Ohm termination this setup does not allow crossing point and rise/fall time tests Jitter tests will have only small difference to open circuit test Using a differential active probe provides the open circuit, but can not measure the clock crossing point. Best would be to use two single ended active probes. But there is no GND Pin at the probe connector of the CLB! EKH - EyeKnowHow 3/1/

26 Compliance Test Clock Compliance test require: AC Parametric test Jitter test Available tools PCISig ClockJitter tool: Only Jitter evaluation Scope vendors compliance application Clock Jitter tool (Gen2 Test) Agilent PCIe Gen2 RefClock compliance Test result example Agilent PCIe Gen1 RefClock compliance Test result example EKH - EyeKnowHow 3/1/

27 Compliance Test Clocking/RX architectures Data lane compliance testing utilizes Dual Port methodology Clock and Data are captured with a single acquisition and Sigtest convolves the data to generate a Dataeye. Dual Port Compliance test Methodology does not fit to Data Clocked architecture Data lane compliance does show problems on Clock and Data, but does not allow to distinguish the source of the problem. With the Dual Port test setup in PCIe Gen2 compliance tests it is difficult to analyze the root cause of compliance issues! EKH - EyeKnowHow 3/1/

28 Compliance Test Pass/Fail Clock Compliance Signal Compliance from a System with FAIL Clock Compliance Signal Compliance from a System with PASS Clock Compliance EKH - EyeKnowHow 3/1/

29 CLK Compliance Test Pass/Fail Clock Compliance Failing Clock Compliance Passing Clock Compliance 5ps vs. 2ps RMS jitter EKH - EyeKnowHow 3/1/

30 CLK Clock Compliance Test: Post Processing for Gen2 The clock compliance test algorithms requires quite some post processing of the Data Two clocking architectures Clock Filter Functions RefClk RX Architecture Common Clocked Data Clocked Two frequency ranges Jitter >1.5MHz Jitter < 1.5MHz SSC Separation PLL difference function MHz step BPF PLL difference function 1.5MHz step HPF Edge filtering No SSC Separation Max. PLL BW function MHz step BPF Max. PLL BW function 1.5MHz step HPF Edge filtering - SSC removal - Transfer to freq-domain - BP and HP Filters - PLL Transfer function manipulation - Edge filter to clean up scope sampling risidual For failure analysis these functions need to be rebuild outside of the scopes compliance application EKH - EyeKnowHow 3/1/

31 CLK Compliance Test Pass/Fail Clock Compliance PASS CLK TIE Fail CLK TIE EKH - EyeKnowHow 3/1/

32 CLK Compliance Test Pass/Fail Clock Compliance Combined Jitter LF jitter HF jitter LF Jitter <1.5MHz HF Jitter > 1.5MHz How to create this data: - SSC removal - Transfer to Frequency domain - BP and HP Filters - PLL Transfer function manipulation HP Filter from 1.5MHz to 50MHz BP Filter from 0.1 to 1.5MHz EKH - EyeKnowHow 3/1/

33 CLK Compliance Test Pass/Fail Clock Compliance RX/TX PLL transfer functions for RMS jitter calculation for common clock RX Architecture These are spec defined PLL transfer functions to calculate RMS jitter (not measured curves) EKH - EyeKnowHow 3/1/

34 CLK Compliance Test Failure analysis Based on the information from the clock compliance test the conclusion is that the bad signal quality on the data is caused by a clock jitter problem. To do more detailed analysis the compliance tests need to be re-build in the normal scope environment. As some post-processing features (e. g. PLL difference function) is a special feature of the compliance application this might be difficult. Functions that should be available on most scopes directly: Clock recovery by second order PLL SSC removal (better to switch off SSC for analyis) FFT for transfer of TimeDomain signal to Frequency Domain BandPass and HighPass Filters TIE measurement for Clock jitter EKH - EyeKnowHow 3/1/

35 CLK Compliance Test Failure analysis TIE shows ~3MHz Jitter on Clock with standard Scope tool functionality 10 Cycles measured EKH - EyeKnowHow 3/1/

36 CLK Compliance Test Failure analysis Jitter is too fast for Number 1 Jitter source: DCDC switching Power Noise 3MHz noise might already come out from the host clock generator and it might be not be possible to Improve on board level Possible solution (?): Turn drawback from embedded systems into advantage Zero Delay clock buffers can clean up input clock by PLL Bandwidth setting optimization Always configure the Zero Delay clock buffer according to your needs! Measure input clock compliance as well! EKH - EyeKnowHow 3/1/

37 Agenda 1) PCI-Express Clocking 2) Motivation and Background 2a) Basics 2b) Clocking in different PCIe generations 2c) Different PCIe clocking architectures 3) Clock Compliance Test 4) Conclusion EKH - EyeKnowHow 3/1/

38 Conclusion Clock and Data Compliance Test on embedded Systems with PCIe interface is required and really important! Don t forget TX AND RX for Data as well! 100MHz is NOT low speed On clocks Jitter is important! Edge rates are important for signal quality The standard test setup with the CLB does not allow to do all measurements with a single setup Single ended open circuit measurements for e. g. crossing point tests are difficult to implement Especially for embedded applications the transport delay delta for common clocked architectures need to be considered EKH - EyeKnowHow 3/1/

39 Conclusion The convolution of Data and Clock signals in the Dual Port Test setup for Gen 2 TX compliance tests makes it difficult to distinguish between clock and data lane related issues. Most systems will require compliance to both, common and data clocked RX architecture Cascading PLLs can cause issues due to jitter amplification, but might be also used to fix clock jitter issues For analysis of Clock jitter issues it is required to understand the post processing of the scopes compliance application. EKH - EyeKnowHow 3/1/

40 References [1] PCI-Sig, PCI Express Jitter Modeling, Revision 1.0, July 2014, p. 19. [2] Agilent Technologies, N5393C PCI Express 3.0 (Gen3) Software for Infiniium Oscilloscopes, EN, May 2013, p. 15. [3] Agilent Technologies, Jitter Analysis Techniques for High Data Rates, cp.literature.agilent.com/litweb/pdf/ en.pdf, February 2003, p. 4. [4] Ernie Buterbaugh, Cypress, " Cypres PerfectTiming II, Perfect Timing II: Design Guide for Clock Generation and Distribution, Chapter 3-3. EKH - EyeKnowHow 3/1/